TSMC Reportedly to Manufacture SoftBank's AI Chips, Replacing Intel

[Nomad76]

SoftBank has reportedly decided against using Intel's foundry for its ambitious AI venture, Project Izanagi, and is opting for TSMC instead. The conglomerate aims to challenge NVIDIA in the AI accelerator market by developing its own AI processors. This decision marks another setback for Intel, which has faced several challenges recently. In February 2024, reports emerged that SoftBank's CEO, Masayoshi Son, planned to invest up to $100 billion to create a company similar to NVIDIA, focused on selling AI accelerators. Although SoftBank initially worked with Intel, it recently switched to TSMC, citing concerns about Intel's ability to meet demands for "volume and speed."

The decision, reported by the Financial Times, raises questions about Intel's future involvement and how SoftBank's ownership of Arm Holdings will factor into the project. While TSMC is now SoftBank's choice, the foundry is already operating at full capacity, making it uncertain how it will accommodate this new venture. Neither SoftBank, Intel nor TSMC has commented on the situation, but given the complexities involved, it will likely take time for this plan to materialize. SoftBank will need to replicate NVIDIA's entire ecosystem, from chip design to data centers and a software stack rivaling CUDA, a bold and ambitious goal.





In July, SoftBank expanded its semiconductor portfolio by acquiring Graphcore, a British AI chip designer. While the acquisition amount remains undisclosed, this move is consistent with SoftBank's significant presence in the chip industry. The company already holds a majority stake in Arm, another British chip designer, which it purchased for $32 billion in 2016. Despite Arm's return to the stock market last year, SoftBank maintained its controlling interest.

In a separate development, Intel divested its position in Arm. The American tech giant sold its 1.18 million shares, generating approximately $146.7 million from the transaction.

View at TechPowerUp Main Site | Source

 

[TumbleGeorge]

Too many strikes on Intel. May can try ask for bankruptcy?

 

[Kn0xxPT]

Too many strikes on Intel. May can try ask for bankruptcy?

Well ... probably not yet ... but i suspect that Nvidia is keeping an eye on Intel "buy" options ... or CPU co-lab.

 

[Frank_100]

Too many strikes on Intel. May can try ask for bankruptcy?

Intel is going to be just fine.

They were last to move to EUV and have suffered.

They are going to be the first to move to high NA EUV.
They will have a 1-2 year advantage on everybody.

Now if they would start making chips for the HEDT market (16 P-cores, avx512).

 

[tfp]

Too many strikes on Intel. May can try ask for bankruptcy?

For as frequent as AMD was making losses prior to 2016, Intel will take a long time to die.

https://www.nasdaq.com/market-activity/stocks/amd/revenue-eps

 

[john_]

Intel needs to build it's own CPUs also in it's fabs, meaning they need to have extra capacity of good wafers to also cover orders from their customers. Meaning this isn't necessarily bad news for Intel. Their future process nodes might doing just fine, but not being currently at the needed capacity to cover also others.

 

[kondamin]

So Softbank is going to make a processor that is bound for failure at a more expensive plant that is going to be less willing to cater to their whims, and other great move by softbank brilliant.

 

[tfp]

So Softbank is going to make a processor that is bound for failure at a more expensive plant that is going to be less willing to cater to their whims, and other great move by softbank brilliant.

They really should be at Intel if they want to fail, amirite?

 

[Frank_100]

AI costs have to come down.
This will probably mean a new architecture.
In memory computing with phase change memory or ReRam.

 

[kondamin]

They really should be at Intel if they want to fail, amirite?

Intel that has to prove it self in the foundry business for third parties will be far more willing to help them out than overbooked tsmc.

so no while it’s very unlikely SoftBank will get anywhere in its latest foley it’s going to cost them less breaking ground with the help of intel than it will with tsmc

 

[tfp]

Intel that has to prove it self in the foundry business for third parties will be far more willing to help them out than overbooked tsmc.

so no while it’s very unlikely SoftBank will get anywhere in its latest foley it’s going to cost them less breaking ground with the help of intel than it will with tsmc

They could have gone Samsung as well, tsmc must be providing something they think they need.

 

[kondamin]

They could have gone Samsung as well, tsmc must be providing something they think they need.

yes it being the number #1 foundry in the world should go over great with the share holders and future investors.
Tsmc building a fab in Japan would be a second.

Technology wise meeeh, they need a design first

 

[InVasMani]

AI costs have to come down.
This will probably mean a new architecture.
In memory computing with phase change memory or ReRam.

We have started hearing about compression on system memory type technologies emergence which is nice. I mentioned that ages ago that would be really great to see DIMM's integrating a chip to compress the memory storage directly to both help in speeding up data transmission and increase storage capacity. Even a fairly minor bit of that initially and improving it over time should go a long way. I feel there is defiantly room for it to mature. Instead of double dimm's for capacity we might start seeing double sided dimm's, but one side is for capacity on the opposite side are chips that of similar size with TSV that compress/decompress the individual DRAM banks in real time. At least that's my thoughts on possible way to go about it. It seems like a pretty plausible methodology to go about it. The same could be applied with NAND as well.

I don't think GPU's likely be doing that given they already have a GPU core that's perfectly suited for math and additionally the higher capacity I think is both more vital and precious in terms of real estate space than memory dimm's are on a PC system for the average consumer user. In the case of servers perhaps not, but they don't represent standard usage so needs will vary in some cases you might need tons of capacity while in others.

Another angle on the memory thing is maybe they'll license FPGA technology or something and could both be used for compression/decompression at times, but also depending on usage be reconfigured as programmable co-processors integrated into system memory. That would be kind of neat novelty to see happen or maybe they'll license CPU/GPU cores from various manufacturers to create a kind of SOC on the memory dimm's themselves that can do compression/decompression as well as other specialized tasks. That's just me trying to think outside the box though on new ways to fuse and merge some of these technology components together in some new ways. I'm sure there would be some positives and negatives to any of these idea's naturally.

 

[tfp]

At one point Intel and I think AMD had plans to put FPGA on package. The problem with FPGA is they are great for proof of concept but generally they don't seem to be used long term. People move to ASIC, GPUs, or something else.

 

[R0H1T]

They will have a 1-2 year advantage on everybody.

Advantage with machines? Is that why they squandered a 5(?) year lead when they moved to 22nm & everyone was so far behind

 

[ScaLibBDP]

At one point Intel and I think AMD had plans to put FPGA on package. The problem with PFGA is they are great for proof of concept but generally they don't seem to be used long term. People move to ASIC, GPUs, or something else.

>>...At one point Intel and I think AMD had plans to put FPGA on package....

Intel spent a lot of money and did it. A complete "infrastructure", I mean a package with CPU-plus-FPGA and software is Too expensive. It makes sense Only for Data Centers, and unfortunately for Intel, many providers of Data Center solutions use GPUs from NVIDIA and AMD.

 

[Daven]

"...citing concerns about Intel's ability to meet demands for "volume and speed."

Or at least they could do it fast enough at scale as long as Soft Bank was okay with the chips oxidizing and frying due to runaway voltage after one year of operation.

 

[ScaLibBDP]

>>...SoftBank will need to replicate NVIDIA's entire ecosystem, from chip design to data centers and a software
>>stack rivaling CUDA, a bold and ambitious goal...

$100 Billion is Not enough. I'm sure for 99.99% that it will never happen.

 

[remixedcat]

Intel that has to prove it self in the foundry business for third parties will be far more willing to help them out than overbooked tsmc.

so no while it’s very unlikely SoftBank will get anywhere in its latest foley it’s going to cost them less breaking ground with the help of intel than it will with tsmc

w the raptorlakegate issue rn ppl aren't gonna trust intel fabs

 

[tfp]

>>...At one point Intel and I think AMD had plans to put FPGA on package....

Intel spent a lot of money and did it. A complete "infrastructure", I mean a package with CPU-plus-FPGA and software is Too expensive. It makes sense Only for Data Centers, and unfortunately for Intel, many providers of Data Center solutions use GPUs from NVIDIA and AMD.

Yeah looks like as far as AMD has gotten is putting an PFGA on a x86 board and is still targeting embedded.

AMD’s x86 CPU and FPGA tango on Sapphire Technology’s Embedded+ PC motherboard

What could possibly be interesting about a new PC motherboard? New x86 CPU? Yawn. Fancy new cooling fan with blue LEDs? Oh puh-leeze! Integrated FPGA for hardware I/O and computational acceleration…

www.eejournal.com

 

[Wirko]

We have started hearing about compression on system memory type technologies emergence which is nice. I mentioned that ages ago that would be really great to see DIMM's integrating a chip to compress the memory storage directly to both help in speeding up data transmission and increase storage capacity. Even a fairly minor bit of that initially and improving it over time should go a long way. I feel there is defiantly room for it to mature. Instead of double dimm's for capacity we might start seeing double sided dimm's, but one side is for capacity on the opposite side are chips that of similar size with TSV that compress/decompress the individual DRAM banks in real time. At least that's my thoughts on possible way to go about it. It seems like a pretty plausible methodology to go about it. The same could be applied with NAND as well.

I don't think GPU's likely be doing that given they already have a GPU core that's perfectly suited for math and additionally the higher capacity I think is both more vital and precious in terms of real estate space than memory dimm's are on a PC system for the average consumer user. In the case of servers perhaps not, but they don't represent standard usage so needs will vary in some cases you might need tons of capacity while in others.

Another angle on the memory thing is maybe they'll license FPGA technology or something and could both be used for compression/decompression at times, but also depending on usage be reconfigured as programmable co-processors integrated into system memory. That would be kind of neat novelty to see happen or maybe they'll license CPU/GPU cores from various manufacturers to create a kind of SOC on the memory dimm's themselves that can do compression/decompression as well as other specialized tasks. That's just me trying to think outside the box though on new ways to fuse and merge some of these technology components together in some new ways. I'm sure there would be some positives and negatives to any of these idea's naturally.

Doesn't seem logical. You'd have to transfer uncompressed data, which means more data, over the memory bus. A better example of processing in memory would be an algorithm to search for certain patterns in memory. In this case, only the searched part would have to be sent over the bus to the processor, instead of everything.

 

[ScaLibBDP]

Doesn't seem logical. You'd have to transfer uncompressed data, which means more data, over the memory bus. A better example of processing in memory would be an algorithm to search for certain patterns in memory. In this case, only the searched part would have to be sent over the bus to the processor, instead of everything.

>>...Doesn't seem logical...

Why? On Windows platforms compression of files is a very-very old and mature technology! Next, starting from Windows 10 there is a Memory Compression feature and take a look at a Task Manager ( memory related attributes ). So, from software side Microsoft has implemented it Twice . The only problem is that a hardware based solution for Memory Compression will be more complex. For example, all external Memory Extenders are Not for a regular PC consumer because all these Memory Extenders are very expensive. Take a look at prices for CXL Memory Expander by Samsung.

 

[Wirko]

>>...Doesn't seem logical...

Why? On Windows platforms compression of files is a very-very old and mature technology! Next, starting from Windows 10 there is a Memory Compression feature and take a look at a Task Manager ( memory related attributes ). So, from software side Microsoft has implemented it Twice . The only problem is that a hardware based solution for Memory Compression will be more complex. For example, all external Memory Extenders are Not for a regular PC consumer because all these Memory Extenders are very expensive. Take a look at prices for CXL Memory Expander by Samsung.

I'm not saying ompression in general is a bad idea. But if you're doing it on the memory modules, you need to transfer uncompressed data over the memory bus, and that may be a serious bottleneck. If the CPU (or whatever sort of processor) is doing it, it sends and receives already compressed data to/from memory.
Remember DirectStorage and GPU decompression? One of its goals was (is?) to avoid sending uncompressed data to the GPU over (relatively) slow PCIe bus.
There's at least one more issue: the compression ratio is unpredictable. How would a smart DIMM handle that?

 

[DaemonForce]

Intel is going to be just fine.

Hahahahaah ✖
That's a black swan moment.
Intel is done.
SoftBank and TSMC wasn't in my bingo chart this year and that's some dark magic.

>>...SoftBank will need to replicate NVIDIA's entire ecosystem, from chip design to data centers and a software
>>stack rivaling CUDA, a bold and ambitious goal...

It's also SoftBank. They find a way.

 

[R-T-B]

Hahahahaah ✖
That's a black swan moment.
Intel is done.
SoftBank and TSMC wasn't in my bingo chart this year and that's some dark magic.

He's not wrong. Intel is NOT done. They haven't even spun off their fab arm for a quick cash influx yet like AMD already did many years ago, for example. They are hurting right now, but awash with options to sell and I can promise you no one is discussing bankruptcy outside enthusiast fanboys.